Feature extraction concepts#

Feature extraction combines an image with a region-of-interest (ROI) mask. The morphological mask defines its shape; the intensity mask contains the voxel values used for intensity and texture features. The guides for Re-segmentation guidelines and Discretization guidelines explain how to prepare that intensity population.

The concepts below apply to both the GUI and Python API. See Understanding results for feature names and output metadata.

Choose texture aggregation#

The dimension controls whether texture neighbourhoods stay within slices or extend through the volume:

Dimension	Texture calculation	How slices are combined
`2D`	Calculate texture within each slice.	Combine the resulting feature values across slices.
`2.5D`	Calculate texture within each slice.	Merge matrices across slices before calculating features.
`3D`	Calculate texture across the volume, including between slices.	Use the volume’s texture matrices.

For directional features such as the grey level co-occurrence matrix (GLCM) and grey level run length matrix (GLRLM), averaged calculates features for each direction and averages the values; merged combines matrices before calculating features. 2D, slice-merged merges directions within each slice, while 2.5D, direction-merged merges slices for each direction. Other texture families use their own dimension-specific aggregation rules.

For 2D extraction, Slice Averaging offers Mean, Weighted Mean (weighted by ROI voxel count), and Median. Keep the dimension, aggregation, and slice-averaging settings consistent across cases and record them with the results.

GUI and Python aggregation settings#

The Python column names below are arguments to Radiomics. Batch extraction uses aggregation_dimension and aggregation_method for the same values.

GUI selection	`aggr_dim`	`aggr_method`
`2D, averaged`	`"2D"`	`"AVER"`
`2D, slice-merged`	`"2D"`	`"SLICE_MERG"`
`2.5D, direction-merged`	`"2.5D"`	`"DIR_MERG"`
`2.5D, merged`	`"2.5D"`	`"MERG"`
`3D, averaged`	`"3D"`	`"AVER"`
`3D, merged`	`"3D"`	`"MERG"`

For 2D slice averaging, the Python defaults select the mean. Set slice_weighting=True for the voxel-weighted mean or slice_median=True for the median; these options are mutually exclusive.

Feature families#

The available families depend on the image dimensionality and prepared ROI data. GUI and batch extraction select the supported families automatically; in the single-ROI Python API, use families or features to select them:

morphology
local intensity
intensity statistics
intensity histogram
intensity-volume histogram (IVH)
grey level co-occurrence matrix (GLCM)
grey level run length matrix (GLRLM)
grey level size zone matrix (GLSZM)
grey level distance zone matrix (GLDZM)
neighbourhood grey tone difference matrix (NGTDM)
neighbouring grey level dependence matrix (NGLDM)

For the preparation required by each family, see Python image workflows and Discretization guidelines. Morphology requires a 3D ROI. See Radiomics for the full API.

ROI size requirements#

For volumetric images, Z-Rad validates the morphological mask for the requested feature families. Texture analysis uses the selected aggregation dimension:

For 3D extraction, the mask must contain at least 27 valid voxels, and the bounding box of the nonzero mask region must be at least 3 voxels wide in every dimension.
For 2D and 2.5D extraction, Z-Rad validates each slice independently. A slice is discarded if it contains fewer than 9 valid voxels or if its nonzero bounding box is smaller than 3 voxels in either in-plane dimension.
If no slice satisfies these 2D or 2.5D requirements, radiomics extraction is aborted for that mask.

Morphology and other families that use a volumetric ROI retain their 3D validation rules even when texture aggregation is slice-wise. Single-slice images follow a separate 2D extraction path. Re-segmentation can further reduce the voxels available to intensity-based features; an empty intensity ROI cannot be used for those calculations. See Troubleshooting for rejected masks.